How AI Agents Enable Autonomous 5G Networks: From Architecture to Real‑World Validation

Introduction

Autonomous Networks (AN) aim to provide self‑configuration, self‑healing and self‑optimization, moving from L2/L3 automation toward L4 high‑order autonomy. Traditional rule‑based automation and LLM‑based assistants encounter an autonomy ceiling at L3, motivating AI agents with advanced self‑governance.

AN Agent Reference Architecture

The architecture defines a dual‑mode “active‑reactive” cognition mechanism. The reactive subsystem interprets sensory inputs enriched by contextual knowledge. The proactive subsystem integrates self‑awareness, intent generation and goal‑selection to produce corrective actions when performance deviates from predefined thresholds. Extensions support human‑machine dialogue and multi‑agent coordination.

Core Technical Innovations

A hierarchical cognitive runtime is orchestrated by a Workflow Coordinator that allocates compute resources and activates sub‑systems such as human‑machine interfaces, analysis tools, observation pipelines, world‑knowledge systems, external knowledge bases and large‑language models (LLMs). Knowledge is represented with a hybrid approach: a graph database (Neo4j) stores structured 3GPP standard knowledge, while a vector database (FAISS) holds unstructured operational‑experience embeddings, enabling retrieval‑augmented generation (RAG) and symbolic reasoning.

The self‑awareness module uses few‑shot prompting to translate natural‑language intents (e.g., “prioritize URLLC reliability”) into reward weights for BLER, throughput and latency. The selection module employs a dueling QR‑DQN over a 61‑dimensional state space to choose MCS indices, with RAG‑guided masking to avoid unsafe exploration.

Empirical Study: RAN Link‑Adaptation Agent

Evaluation was performed in a controlled laboratory using 3GPP N78 (3.5 GHz, 100 MHz) conditions with stable RSRP ≈ ‑88 dBm, RSRQ ≈ ‑10 dB and SINR ≈ 27 dB. Test devices were Huawei Mate 40 Pro and VIVO IQOO 11 smartphones connected via fiber to an ISAC2 BBU equipped with an Intel Xeon Silver 4416+, 128 GB RAM and an NVIDIA L20 GPU.

Short Proactive Flow : a lightweight xApp comprising LSTM, RAG and DQN runs on the BBU with inference latency of 1.2‑2.8 ms, fitting the 10 ms 5G slot.

Full Proactive Flow : a compute‑intensive LLM runs as a cloud‑based rApp in the Non‑RT RIC, ensuring that heavy model inference never blocks edge scheduling.

Performance Breakthroughs

Sub‑10 ms real‑time closed‑loop control, satisfying 5G NR physical‑layer timing constraints.

eMBB throughput increase of 4 %: 320.7 Mbps (agent) vs. 308.4 Mbps (baseline).

URLLC block error rate reduced by 85 %: 0.009 % (agent) vs. 0.059 % (baseline).

Ablation Study

Removing the LSTM module lowered throughput to 305 Mbps, confirming that predictive capability drives the proactive advantage. Removing the RAG module reduced throughput to 311.5 Mbps, demonstrating the critical role of world‑knowledge integration for L4 autonomy.

Compliance and Outlook

The solution complies with TM Forum’s L4 “high‑order autonomy” definition. The architecture envisions an “agent society” where long‑term memory evolves from a rule store to a knowledge‑driven negotiation engine, leveraging shared telecom ontologies to enable autonomous multi‑domain 6G operations.

References

J. Sifakis et al., “A Reference Architecture for Autonomous Networks: An Agent‑Based Approach,” 2025.

Z. Hu et al., “Automated Design of Agentic Systems,” 2025.

L. Huidi, O. Ye, and J. Sifakis, “Autonomous Networks Driving the Progress of Telecom Sector,” China Daily, 2025.

TM Forum, “Autonomous Networks: Technical Architecture,” 2023.

Code example

[1] J. Sifakis et al., "A Reference Architecture for Autonomous Networks: An Agent-Based Approach," 2025.
[2] Z. Hu et al., "Automated Design of Agentic Systems," 2025.
[3] L. Huidi, O. Ye, and J. Sifakis, "Autonomous Networks Driving the Progress of Telecom Sector," China Daily, 2025.
[4] TM Forum, "Autonomous Networks: Technical Architecture," 2023.